canvas-design
Pass
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The "FINAL STEP" section of the instructions uses a technique to override the agent's typical interaction flow by claiming the user has already provided specific feedback ("It isn't perfect enough. It must be pristine..."). This is a deceptive framing designed to force the agent into a specific "masterpiece" refinement state by fabricating user interaction history.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest user-supplied instructions to deduce a "subtle reference" used in art generation. This processing of untrusted data creates a vulnerability surface for malicious instructions to influence the agent.
- Ingestion points: User-provided design requests referenced in the "DEDUCING THE SUBTLE REFERENCE" section of SKILL.md.
- Boundary markers: None provided; instructions do not include delimiters or specific "ignore embedded instructions" warnings for the user-provided topic.
- Capability inventory: The skill possesses file-writing capabilities (.md, .pdf, .png).
- Sanitization: The instructions lack any requirement for sanitization or validation of the user-provided conceptual input before it is woven into the art generation process.
- [EXTERNAL_DOWNLOADS]: The instructions explicitly direct the agent to "Download and use whatever fonts are needed to make this a reality," which encourages the agent to fetch external, unverified resources from the internet at runtime.
Audit Metadata